Principle:Google deepmind Mujoco Control Noise Generation
| Knowledge Sources | |
|---|---|
| Domains | Benchmarking, Stochastic_Processes, Control |
| Last Updated | 2026-02-15 06:00 GMT |
Overview
Pattern for generating temporally correlated control signals using an Ornstein-Uhlenbeck process for realistic simulation benchmarking.
Description
Control Noise Generation produces smooth, temporally correlated actuator commands for benchmarking purposes. Rather than using independent random noise at each step (which creates unrealistically jerky motion), an Ornstein-Uhlenbeck (OU) process generates signals that mean-revert with controllable rate and amplitude. This results in more realistic actuation patterns that exercise the full dynamics of the model including contacts.
The Halton quasi-random sequence is used instead of pseudo-random numbers to ensure reproducible and evenly distributed noise.
Usage
Use in benchmarking scripts to generate realistic control inputs that exercise the full simulation pipeline including contacts and constraint solving.
Theoretical Basis
The Ornstein-Uhlenbeck process is a mean-reverting stochastic process:
In the discrete MuJoCo implementation:
- corresponds to ctrl_noise_rate
- corresponds to ctrl_noise_std
- uses Halton sequence values mapped through the inverse normal CDF